Professional Documents
Culture Documents
Common: A I, I E. 1, Some Set
Common: A I, I E. 1, Some Set
L.M. BREGMAN
Leningrad
(Received 20 Pay 1966)
Let us consider the function 11(x, y), defined over 5’ x S, and satis-
fying the following conditions.
200
Finding the common point of conver sets 201
This point x ;.ill be called the D-projection of the point y onto the
set Ai.
Lemmaf
Let us now consider soae relaxation controls for wtlich the limiting
point of the sequence {,~n) belongs to the intersection of the sets ‘1i,
y indirbq tb camnon point of convex sets 203
Let i = i:, 2, ,.., d and let tbe indices be chosen in cyclic order,
i.e. iO(sO) = 1, ii = 2,. . . , &,+-f(zf+-i) = m, i,(xa) = 1, and so on.
Tnon any limiting ooint Y* oE the relaxational sequence {xn) is a co%v)n
poink of the JB~S Ai,
’ i.301‘. Let n* ‘0” the limiting point of the sequence {xn) and xnk * Y*.
la..
:b sF,patate out from the sequence, (~‘~1 a subsequence which is wholly
c%kzined in one of the sets Ai, Eor exanple in A,. We shall assu!ne that
{ L"'>Ic .I1. We separate out from the sequences {X nkti-l) those w:-lich
are cmveixz:lt. It can be assuned that the sequences {xnkti-‘) thecl-
se~~v~s are c9svzrgen:. 1,et
Ilk+1
Since according to Lemma 2 we have :I(x , xnk) 4 0, therefore also
(1) The set .s is closed, and for any ~1, z2 E ri 11 ‘j the function
T{(y) = fi(Zl, y) - D(z*, y) is continuous over S;
*= **
x x .
1. Let Y be a real Hilbert space, .i‘ = .I( and iJ(x, ,y) = (X - ,y, x - y).
G(z)=D(z,y)--(z,Piy)=(z-Y,z-Y)-
- (2 - piy, 2 - Piy) = 2(Z,PiY - Y) + (Y, Y) - PiYPPiY)
Furthermore,
Consequently, lim(u, xnk ) = lim(u, yn’) = (u, ,y*), and this means that
C(y)=- (Zj - y, 21 -- y) - (22- Y, zz- y) = (Zi, 21) - (zz,z2) + 2(% -- a, Y)
D(~,~)=f(~)-.f(Y)-(g(~),~---_). (1.4)
G(z) = --f(Y)+ftPid)-(g(y),y)+(g(Pirl).PiEI)-
---(R(~)--gg~iY),z)-
PurtherrnorP-I :;,y ’ c v , y ) = ,5 (Y) - 6 (Y) = 0, so condition IV is satis-
fied.
An example of a function P(X). for which the function C(X, ,y), con-
structed according to formula (1.4). satisfies conditions I - VI and
(l), is given by a positive definite quadratic form given over the
whole space SP. Tine corresponding function !)(n, +y) = (X - y, Ll(x, y)),
as is easy to see, satisfies conditions I - VI end (1).
f(x)= 5 xjhlsj,
j=f
given over the set 5’ = (x E Ep Ix > O}. Its corresponding function
i)fx, y) has the form
Let us verify that function (1.5) satisfies condition VI. Let i,(xn,
yn) + 0 and yn -+ y* = (yl*, .y2*, . . . , ,yp*). If yj* = 0, then also
XI.n -+ 0, since otherwise 3(xn, yn) -+ a. If :fj* > 0, then “in + ,JJ~*in
view of the continuous nature of the function 1(x, y) when yi > 0.
and
to minimize
f(x) (2.1)
subject to the conditions
Ax=b, (2.2)
x E s. (2.3)
iiere A = 11 aij 11 is a matrix with IR rows and p columns, x E ,!?p, b E Em.
Let us denote the i-th row of the matrix A by Ai. We assume that all
Ai # 0.
(2.6)
- @?(Y”W - y”) - II (x8, y”) = D (y’, y”) - D p*, Y”) .
Theorem .jr
Proof. Let xn+l be a D-projection of the point xn onto the set idi.
Then
Note 2. Both xn+I and A are uniquely determined from the conditions
(2.7) and (2.8).
Then
Minimize
f(x) (2.11)
Ax > 3, (2.12)
x E 6.9. (2.13)
Let the function f(n) have the same properties as earlier, and let
R= {xt:E~(Ax>b,x~S} #A.
Let ,z, = {x E SI, there exists u = (uI, u2, . . ., urn) such that u>O.
3. (a) If (A+, xn) <b%, then for x ntI we take the D-projection
of the point nn on the set Ai, = {ZE .&‘*I(At,., x) = b,,,}, i.e. nntl is
(c) If (A,,, P) > bi, and ui,,> 0, we determine xntl from the
relationship
2 12 L.M. Bregman
n+i = n+i
:and un+l is determined from the formula u.
In
Uin
n pm ui = ui when
i f in.
Here
and n,,’ = ui n.
n
Theorem 4
3. Let z CF R. Then
zl(z,5~j=f(Z)-j+~)-(g(~~),Z-X~)= (2.25)
= f(2) - f(~nj - (u+,z -P) <‘f(s) - f(~j -
- (ZP, b - AZ?) = f(z) - cp (xn, un) < f(a) - q (x0, u”).
(2.261
q WY @j Q f(z),
This together with (2.22) shows that there exists
(2.27)
any limiting point x* of the sequence {xn) belongs to the set X. Apart
from this, since condition (2) is satisfied, we have xn -. x* E R.
5. Let
We choose N so that when Q > IVfor i E 11, we have (Ai., xn) > Oi.
Then when n > N + m for i E lr, we obtain ui” = 0. Therefore when
n>N+m
=(g(zn),z~-z*)=D(2*,zy--f(5*)+f(Z~).
to maximize
0) = u-4 (2.29)
a >o, x ES. ww
From Theorem 4 we have min f(x) = sup q~(n, y), where the minimumis
taken over the set of vectors n, satisfying conditions (2.12) and (2.13).
while the upper bound is taken over the whole set of pairs (x, u),
satisfying conditions (2.29) and (2.30).
5. If the sequence {an) has the limiting point IL*, the pair (x*,u*)
is a solution of the problem (2.28) - (2.30).
to minimize
Finding the common point of convex sets 215
i Xj III Xj (2.31)
j=i
“j = Yj exp(hj) , (2.34)
where h is the unique root of the equation
P
$ =
Xij Ui, $ Xij = bj.
j=i i=i
Finally we remark that the relaxation method can be used for the
approximate solution of problems in linear progremning.
to minimize
i CjXj (2.36)
j-i
subject to the conditions (2.37) and (2.38). Here E > 0, and f(x) is.
the same as earlier.
If for f(n) we take the function 2 xjh% we can solve the prob-
j-1
lem of minimizing the function (2.39) aith only the restriction (2.37),
since the restrictionsnj > 0 will be automatically satisfied.
Translated by G. Kiss
Finding the common point of convex sets 2 17
REFERENCES
2. MOTZKIN, T.S. and SCHOENBERG, 1.1. The relaxation method for linear
inequalities. Can. J. #ath. 6, 3, 393 - 404, 1954.
4. BREGHAN, L.M. Finding the common point of convex sets by the method
of successive projections. Dokl. Akad. Xauk SSSR 162, 487 - 490,
1965.
7. . DoRN, W.D. A duality theorem for convex programms. IBM J. Res. Dew.
4, 4, 407 - 413, 1960.